Hook callbacks¶

This provides both a standalone class and a callback for registering and automatically deregistering PyTorch hooks, along with some pre-defined hooks. Hooks can be attached to any nn.Module, for either the forward or the backward pass.

We'll start by looking at the pre-defined hook ActivationStats, then we'll see how to create our own.

ActivationStats saves the layer activations in self.stats for all modules passed to it. By default it will save activations for all modules. For instance:

The saved stats is a FloatTensor of shape (2,num_modules,num_batches). The first axis is (mean,stdev).

So this shows the standard deviation (axis0==1) of 2th last layer (axis1==-2) for each batch (axis2):

Internal implementation¶

Callback methods¶

You don't call these yourself - they're called by fastai's Callback system automatically to enable the class's functionality.

Registers and manually deregisters a PyTorch hook. Your hook_func will be called automatically when forward/backward (depending on is_forward) for your module m is run, and the result of that function is placed in self.stored.

Deregister the hook, if not called already.

Acts as a Collection (i.e. len(hooks) and hooks[i]) and an Iterator (i.e. for hook in hooks) of a group of hooks, one for each module in ms, with the ability to remove all as a group. Use stored to get all hook results. hook_func and is_forward behavior is the same as Hook. See the source code for HookCallback for a simple example.

Deregister all hooks created by this class, if not previously called.

Convenience functions for hooks¶

Function that creates a Hook for module that simply stores the output of the layer.

Function that creates a Hook for all passed modules that simply stores the output of the layers. For example, the (slightly simplified) source code of model_sizes is:

def model_sizes(m, size):
    x = m(torch.zeros(1, in_channels(m), *size))
    return [o.stored.shape for o in hook_outputs(m)]

This method only works on a Learner object with train_ds in it. If it was created as a result of load_learner, there is no data to run through the model and therefore it's not possible to create such summary.

A sample summary looks like:

======================================================================
Layer (type)         Output Shape         Param #    Trainable 
======================================================================
Conv2d               [64, 176, 176]       9,408      False     
______________________________________________________________________
BatchNorm2d          [64, 176, 176]       128        True      
______________________________________________________________________
ReLU                 [64, 176, 176]       0          False     
______________________________________________________________________
MaxPool2d            [64, 88, 88]         0          False     
______________________________________________________________________
Conv2d               [64, 88, 88]         36,864     False     
...

Column definition:

1. Layer (type) is the name of the corresponding nn.Module.

2. Output Shape is the shape of the output of the corresponding layer (minus the batch dimension, which is always the same and has no impact on the model params).

3. Param # is the number of weights (and optionally bias), and it will vary for each layer.

   The number of params is calculated differently for each layer type. Here is how it's calculated for some of the most common layer types:

   • Conv: kernel_size*kernel_size*ch_in*ch_out
   • Linear: (n_in+bias) * n_out
   • Batchnorm: 2 * n_out
   • Embeddings: n_embed * emb_sz

4. Trainable indicates whether a layer is trainable or not.

   • Layers with 0 parameters are always Untrainable (e.g., ReLU and MaxPool2d).
   • Other layers are either Trainable or not, usually depending on whether they are frozen or not. See Discriminative layer training.

To better understand this summary it helps to also execute learn.model and correlate the two outputs.

Example:

Let's feed to a Learner a dataset of 3-channel images size 352x352 and look at the model and its summary:

data.train_ds[0][0].data.shape
learn = cnn_learner(data, models.resnet34, ...)
print(learn.model)
print(learn.summary())

Here are the outputs with everything but the relevant to the example lines removed:

torch.Size([3, 352, 352])

    [...]
    (0): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
    (1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    [...]
    (3): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
    [...]
    (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (8): Linear(in_features=512, out_features=37, bias=True)


======================================================================
Layer (type)         Output Shape         Param #    Trainable 
======================================================================
Conv2d               [64, 176, 176]       9,408      False     
______________________________________________________________________
BatchNorm2d          [64, 176, 176]       128        True      
______________________________________________________________________
[...]
MaxPool2d            [64, 88, 88]         0          False    
______________________________________________________________________
Conv2d               [64, 88, 88]         36,864     False     
[...]
______________________________________________________________________
Linear               [37]                 18,981     True

So let's calculate some params:

For the Conv2d layers, multiply the first 4 numbers from the corresponding layer definition:

Conv2d(3, 64, kernel_size=(7, 7), ...)

3*64*7*7 = 9,408

Conv2d(64, 64, kernel_size=(3, 3), ...)

64*64*3*3 = 36,864

For the BatchNorm2d layer, multiply the first number by 2:

BatchNorm2d(64, ...)
64*2 = 128

For Linear we multiply the first 2 and include the bias if it's True:

Linear(in_features=512, out_features=37, bias=True)

(512+1)*37 = 18,981

Now let's calculate some output shapes:

We started with 3x352x352 image and run it through this layer:

Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)

How did we get: [64, 176, 176]

The number of output channels is 64, that's the first dimension in the number above. And then our image of 352x352 got convolved into 176x176 because of stride 2x2 (352/2).

Then we had:

MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)

which reduced [64, 176, 176] to [64, 88, 88] again because of stride 2.

And so on, finishing with:

Linear(in_features=512, out_features=37, bias=True)

which reduced everything to just [37].

It can be useful to get the size of each layer of a model (e.g. for printing a summary, or for generating cross-connections for a DynamicUnet), however they depend on the size of the input. This function calculates the layer sizes by passing in a minimal tensor of size.

For all modules, uses a callback to automatically register a method self.hook (that you must define in an inherited class) as a hook. This method must have the signature:

def hook(self, m:Model, input:Tensors, output:Tensors)

If do_remove then the hook is automatically deregistered at the end of training. See ActivationStats for a simple example of inheriting from this class.

Callback methods¶

You don't call these yourself - they're called by fastai's Callback system automatically to enable the class's functionality.

Undocumented Methods - Methods moved below this line will intentionally be hidden¶

New Methods - Please document or move to the undocumented section¶